L’EMbeDS DS^3: "Hierarchical Gaussian Processes for Bayesian Modeling and Causal Inference"

The DS³ organizing group, alongside the L'EMbeDS Department of Excellence of the Sant'Anna School for Advanced Studies, is launching a new online seminars series devoted to frontier research in data science, its applications and its implications across disciplines.  

The L'EMbeDS Data Science Seminar Series (L'EMbeDS DS³) hosts national and international scholars to discuss cutting-edge methodology, applications in economics, the social sciences -- and beyond, societal implications and governance issues. 

For announcements and further information please visit our web page and join our Google Group. 

The fourth seminar will feature Samuel Baugh (Penn State University), who will present a talk entitled: “Hierarchical Gaussian Processes for Bayesian Modeling and Causal Inference” 

>>Join the event via the following link 

ABSTRACT:  

While the physical science of climate change provides valuable understanding of the underlying dynamics, inferring these dynamics from observations remains challenging due to the paradoxical "small data" nature of the problem: while the number and representativeness of high-quality climate observations continues to increase, in a sense there will only ever be a single observation of the global climate system as a whole. Inference on the basis of a single observation is a common setting in spatio-temporal statistics, where Gaussian process models have proven to be a flexible and powerful framework for modeling physical and non-physical processes alike. However, despite the advantages of the Gaussian process framework, applications to climate data remain methodologically challenging due to the presence of complex correlation structures (which exhibit extensive non-stationary and anisotropic behavior) as well as the large number of available observations (precluding the direct calculation of likelihoods). In this talk, we discuss the development of computationally feasible hierarchical Gaussian process models for addressing two important inference problems in climate science: constraining equilibrium climate sensitivity from ocean heat content observations, and estimating the causal effect of changing distributions of temperatures on drought conditions. These developments permit observationally-driven estimates of key climate parameters with rigorous uncertainty quantification, which can increase confidence in estimates based on climate simulations. Extensions for integrating observations and climate model simulation output in a unified Bayesian model will be discussed.